0
$\begingroup$

I'm trying to match the prediction interval of the predict.lm() function in R using the formula found in this discussion :

Obtaining a formula for prediction limits in a linear model

I'm using a student's quantile in my interval but in the end it's far larger from the one given by predict().

Is there any specific calculation in the predict function, I tried to look at the code but couldn't find any answer. The formula looks ok as I found exactly the same from others source.

My R code :

airquality_clean <- na.omit(airquality)
attach(airquality_clean)

#Model estimation
model_1 <- lm(Ozone ~., data = airquality_clean)

#Unbias variance of the residuals
sigma_2 <- sum(model_1$residuals**2)/(dim(airquality_clean)[1]-dim(airquality_clean)[2])

#New observation
new <- data.frame(Solar.R=200,Wind=10,Temp=70,Day=1,Month=3)

#Calculated prediction interval
sigma <- sqrt(sigma_2*(1 + as.matrix(new)%*%solve(as.matrix(t(airquality_clean[,-1]))%*%as.matrix(airquality_clean[,-1]))%*%as.matrix(t(new))))
qt <- qt(0.995, df = dim(airquality_clean)[1]-dim(airquality_clean)[2])
int_pred_t <- cbind(predict(model_1, new)-(qt*sigma),predict(model_1, new)+(qt*sigma))
int_pred_t
          [,1]     [,2]
[1,] -22.59931 95.82563

#R prediction interval
predict(model_1, new, interval="predict", level=0.99)}
       fit       lwr      upr
1 36.61316 -21.12916 94.35548

I'm not too far but it's not the same results. If I use a p value from a normal distribution and not a student I'm even closer.

Thank you.

$\endgroup$
4
  • $\begingroup$ Your question is not clear. Are you experiencing a confusion between a confidence interval and a prediction interval? $\endgroup$
    – Glen_b
    Nov 11 '13 at 1:06
  • $\begingroup$ Sorry if i'm not clear.I've tried to numerically calculate the same prediction interval provided by R but I can't match its results. I've used the formula from the discusssion posted in my previous post with the unbiased estimation of σ2 and a treshold from a student distribution.I don't know if there is a problem with the formula or with some numeric aproximations. $\endgroup$
    – Alex
    Nov 12 '13 at 10:26
  • $\begingroup$ It will be a lot easier to figure out what might be wrong if you explain exactly what you did for both parts of the comparison. $\endgroup$
    – Glen_b
    Nov 12 '13 at 10:40
  • $\begingroup$ I've added some code to my previous post. $\endgroup$
    – Alex
    Nov 13 '13 at 13:19
1
$\begingroup$

You didn't construct your new object correctly. You need to take into account the effect of the intercept term by adding a "1" (if you are fitting with an intercept) to your linear function if you are finding the C.I. manually yourself. Check out how I created vector a in the following code. But if you use the predict function to directly find the C.I., then the newdat argument does not need to have any "1" for the intercept. R will take care of that! Check out how I used a.dat below and and found identical results:

> #Model estimation
> lm.1<- lm(Ozone ~., data = airquality_clean)

> #Design Matrix
> x=model.matrix(lm.1)
> 
> #Defining linear function
> a=c(1,200,10,70,1,3)
> 
> #Defining new data.frame
> a.dat=data.frame(Solar.R=a[2],Wind=a[3],Temp=a[4],Month=a[5],Day=a[6])
> 
> #Finding the upper prediction interval
> predict(lm.1,newdat=a.dat)+qt(.995,summary(lm.1)$df[2])*summary(lm.1)$sigma*sqrt(1+t(a)%*%solve(t(x)%*%x)%*%a)
         [,1]
[1,] 103.2434
> 
> #Finding the lower prediction interval
> predict(lm.1,newdat=a.dat)-qt(.995,summary(lm.1)$df[2])*summary(lm.1)$sigma*sqrt(1+t(a)%*%solve(t(x)%*%x)%*%a)
          [,1]
[1,] -16.76174
> 
> 
> #Using predict function
> predict(model_1, a.dat, interval="predict", level=0.99)
       fit       lwr      upr
1 43.24083 -16.76174 103.2434
> 
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.